Platform · Guardrails
Make AI policies enforceable—not just documented.
IntSen Guardrails check governed model calls for prompt injection, personal data, secrets and harmful content, plus rules you write yourself—and record, mask or block according to your settings.
Guardrail verdicts for one turn
- prompt_injectionuser inputblocked · 0.94
- piiuser inputmasked · email
- secretsuser inputpass
- prompt_injectiontool resultpass
- custom: no-customer-contactsreplypass
- harm_textreplyrecorded · low
Illustrative product scenario — synthetic data
The problem
A policy no system enforces is a hope, not a control.
When safety checks live inside individual applications, coverage is uneven, settings drift and nobody can show which policy applied to which call.
- Checks implemented differently—or not at all—by each team
- No evidence of what was detected and what was done
- Tool results returned to the model are rarely checked
Capabilities
What Guardrails does
Input and output checks
Prompt-injection detection
Personal data and secrets
Content safety
Your own rules
Task-level checks
Technical detail
How guardrails are configured and applied
Platform defaults apply to every registration. Owners can switch on stricter settings for a use case or an agent—never weaker ones.
guardrails: promptInjection: { enabled: true, action: block, checkToolResults: true } pii: { enabled: true, action: mask } secrets: { enabled: true, action: block } checkStreamedOutput: during- Actions
- record (shadow), mask (personal data), block
- Where
- user input, tool results, replies, and across the whole turn
- Defaults
- harmful text, prompt injection, personal data and secrets recorded; harmful images blocked at medium severity
- Settings
- in the Registration (use case) and the Agent YAML; dry run shows the effect
- On check failure
- the call continues and the failure is recorded (fail open by default)
- Cost
- model-based checks run on your registration and are shown separately in FinOps
Why it matters
For the people accountable for it
CISO / risk
Consistent controls across every governed agent, with each verdict recorded for investigation.
AI platform
Platform defaults everywhere; teams can only tighten them.
Engineering
Safety settings live in YAML next to the agent, reviewed like code.
Boundaries
What to know before you evaluate
We would rather you hear the limits from us than discover them in a pilot.
- No detection system catches every attack. Guardrails reduce risk; they do not eliminate it.
- Guardrails act on governed model traffic. They do not inspect or block tool executions that happen inside your agent.
- By default a failed or timed-out check lets the call continue (recorded). A fail-closed option is designed; confirm availability for your pilot.
- Automatic user blocking and tool-misuse detection are planned for after the first release.
FAQ
Common questions
Can we test a guardrail before enforcing it?
Yes. Set the action to record: checks run and verdicts are kept, but calls are not blocked. Switch to block or mask when you are confident.
Are the matched contents stored?
Guardrail results store the category and score, never the content that was found.
Can we use our own policy engine?
Yes, through a webhook configured on the agent. IntSen calls it and applies its decision.
Ready to make your enterprise AI accountable?
See how IntSen can help your teams govern access, safety, quality and spending across the AI applications and agents you already run.